⚡ electrical engineering

CAMAL: Improving Attention Alignment and Faithfulness with Segmentation Masks

This paper introduces CAMAL, an efficient and scalable method that leverages segmentation masks as an auxiliary regularizer to significantly improve both the spatial accuracy (alignment) and causal meaningfulness (faithfulness) of attention in vision models across deep learning and deep reinforcement learning paradigms, thereby enhancing explainability without increasing inference costs.

Original authors: Rajdeep Singh Hundal, Yan Xiao, Jin Song Dong, Manuel Rigger

Published 2026-05-12

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Rajdeep Singh Hundal, Yan Xiao, Jin Song Dong, Manuel Rigger

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are teaching a robot to recognize a cat in a photo. You show it thousands of pictures, and eventually, the robot gets really good at saying, "That's a cat!" But here's the problem: why did it think that?

Maybe it's looking at the cat's fluffy ears. Or maybe, it's just looking at the green grass in the background because all the cat photos happened to be taken on lawns. If the robot is just guessing based on the grass, it's not really "learning" about cats; it's taking a shortcut.

This paper introduces a new method called CAMAL (Class Activation Map Attention Learning) to fix this. Think of CAMAL as a strict but helpful teacher who doesn't just check the robot's final answer, but also checks where the robot is looking while it thinks.

Here is how it works, broken down into simple concepts:

1. The Problem: "The Robot is Looking at the Wrong Thing"

In the past, researchers tried to teach robots to look at the right spots using "hints" from other AI models (like a model that knows what a cat sounds like or what it usually looks like). The paper calls these "pseudo-discriminative regions."

The Analogy: Imagine trying to teach a student to find a specific person in a crowd by showing them a blurry, low-resolution sketch of that person. The student might guess the right person, but they might also get confused and point at someone who just happens to be wearing a similar hat. The hint was too vague.

2. The Solution: "The Gold-Standard Map"

The authors noticed that many modern datasets come with segmentation masks.

The Analogy: Instead of a blurry sketch, imagine you have a perfectly traced outline (like a coloring book page) that shows exactly where the cat is, pixel by pixel. This is the "ground truth." It's a human-verified map that says, "The cat is right here, and nowhere else."

CAMAL uses these perfect maps to teach the robot. It says: "When you are looking at the picture, your 'attention' (your mental spotlight) must shine brightly on the areas inside this outline, and stay dark everywhere else."

3. How CAMAL Works (The "Teacher's Penalty")

The paper describes a two-part rule for the robot during training:

Encourage the Good: If the robot's attention shines on the cat (inside the mask), the teacher gives a thumbs up.
Punish the Bad: If the robot's attention shines on the grass or the background (outside the mask), the teacher gives a thumbs down.

The paper calls this "Attention Alignment" (looking at the right spot) and "Attention Faithfulness" (making sure looking at that spot actually helps the robot make the decision). CAMAL forces the robot to do both.

4. The "Batch" Trick (Making it Fast)

Usually, checking where a robot is looking for every single image takes a lot of computer power and time. It's like a teacher stopping to grade every single student's homework one by one before moving to the next.

The Innovation: The authors found a clever math trick to check the whole class at once. Instead of grading 32 students individually, they grade the whole group in one go. This makes the process much faster and allows them to use this method on huge datasets without the computer crashing.

5. What They Found

The researchers tested this on two types of tasks:

Standard Vision (DL): Identifying flowers, pets, and medical images.
Video Games (DRL): Teaching a robot to play a first-person shooter game (ViZDoom).

The Results:

Better Focus: The robots trained with CAMAL looked at the actual objects (the cat, the flower, the enemy) much more accurately than robots trained with the old "blurry sketch" hints or no hints at all.
More Trustworthy: When the researchers tested if the robot's focus actually mattered, they found that CAMAL-trained robots were "faithful." If you covered up the part the robot was looking at, the robot got confused. If you covered up the background, the robot didn't care. This proves the robot was actually looking at the important stuff.
No Speed Penalty: The robots didn't get slower when they were actually playing or identifying things later. The extra work only happened while they were learning.

The Bottom Line

The paper argues that to make AI trustworthy, we need to ground its attention in reliable, human-verified maps (segmentation masks) rather than vague guesses from other AI models. CAMAL is a tool that uses these maps to train AI to look at the right things, making the AI's decisions more logical and less likely to be based on accidental shortcuts.

In short: CAMAL teaches AI to look at the "cat" and not the "grass," using a perfect outline as a guide, and it does so efficiently without slowing down the final result.

Technical Summary: CAMAL

Problem Statement

Vision models deployed in real-world applications require interpretability to ensure their decisions are trustworthy. Class Activation Maps (CAMs) are widely used to visualize a model's attention, highlighting regions the model perceives as discriminative. However, existing approaches suffer from two critical limitations:

Lack of Attention Alignment: Models often attend to spurious regions rather than ground-truth discriminative regions, a phenomenon known as "shortcut learning."
Lack of Attention Faithfulness: Even if attention aligns with a region, it does not guarantee that the attention causally influences the model's decision.

Recent attempts to address this via "attention supervision" have relied on pseudo-discriminative regions derived from pretrained zero-shot models (e.g., CLIP) as reference signals. The authors argue that these priors are untrustworthy because they are optimized for broad generalization rather than precise spatial localization, often resulting in fragmented or spurious reference regions. Furthermore, existing methods often fail to scale efficiently to large datasets because they rely on per-sample gradient-based attention extraction, which incurs prohibitive computational overhead.

Methodology: CAMAL

The authors propose Class Activation Map Attention Learning (CAMAL), a method that leverages segmentation masks as trustworthy reference regions to guide model attention.

1. Trustworthy Reference Regions

Unlike prior works using CLIP-based priors, CAMAL utilizes segmentation masks (often human-verified or generated by tools like SAM) to define ground-truth discriminative regions. These masks provide structured, complete, and accurate spatial information.

2. Mask-Guided Auxiliary Regularization

CAMAL formulates an auxiliary regularization objective that jointly optimizes two metrics derived from the model's attention heatmap $H_c$ and the segmentation mask $M_c$ :

$\alpha$ (In-Mask Attention): The mean attention within the ground-truth discriminative regions.
$\beta$ (Out-of-Mask Attention): The mean attention outside these regions.

The objective function minimizes the difference $(\beta - \alpha)$ , effectively encouraging the model to increase attention within the mask ( $\alpha \uparrow$ ) and suppress attention in spurious regions ( $\beta \downarrow$ ). This is added as a regularizer to the standard cross-entropy loss:
$L = L_{CE} + \lambda \sum_{b} (\beta_b - \alpha_b)$
where $\lambda$ controls the strength of the regularization.

3. Efficient Batch-Level Attention Extraction

A key technical contribution is the resolution of the scalability bottleneck in gradient-based attention extraction (e.g., Grad-CAM). Standard methods require a separate backward pass for each sample in a mini-batch ( $O(B)$ complexity). CAMAL introduces an efficient formulation where the logits for all samples in a mini-batch are summed into a single scalar before performing a single backward pass.

Theoretical Guarantee: Due to the linearity of differentiation and the independence of samples within a batch, this single backward pass yields gradients identical to computing them individually ( $O(1)$ complexity).
Impact: This reduces training overhead significantly, making attention supervision feasible for large-scale training regimes.

4. Architecture and Task Agnosticism

CAMAL is designed to be agnostic to network architecture (applicable to CNNs, Transformers like ViT, and hybrid models) and learning paradigms. In Deep Reinforcement Learning (DRL), where ground-truth labels are absent, CAMAL uses the critic's value estimate $V(s)$ as the target for attention computation, allowing supervision in weakly-supervised settings.

Key Contributions

Principled Attention Supervision: Introduction of CAMAL, which grounds attention learning in trustworthy segmentation masks rather than unverified pseudo-regions.
Novel Regularization Objective: A formulation that simultaneously encourages attention within ground-truth regions and suppresses it elsewhere, addressing the limitations of "suppress-only" regularizers.
Scalable Extraction Method: An efficient batch-level attention extraction technique that reduces computational complexity from $O(B)$ to $O(1)$ , enabling scalable supervision.
Comprehensive Evaluation: A statistically rigorous evaluation across Deep Learning (DL) and Deep Reinforcement Learning (DRL) paradigms.

Experimental Results

The authors evaluated CAMAL against two baselines: Vanilla (no attention supervision) and Prior (attention supervision using CLIP-based pseudo-regions).

Attention Alignment

Metric: Intersection-over-Union (IoU) between model attention and ground-truth masks.
DL Results: CAMAL achieved statistically significant improvements in IoU across all 6 models (ResNet, ConvNeXt, EfficientNet, ViT, Swin, MaxViT) and 3 datasets (BUSI, Oxford-IIIT-Pet, Oxford-Flower). CAMAL's IoU scores were substantially higher than both Vanilla and Prior (e.g., ~33-42% for CAMAL vs. ~2-12% for Prior on various datasets).
DRL Results: In ViZDoom environments, CAMAL showed significantly higher mean IoU compared to Vanilla, with statistical significance confirmed via Stratified Bootstrap Confidence Intervals (SBCI) and Probability of Improvement (POI).

Attention Faithfulness

Metric: Causal impact measured by removing/inserting pixels based on attention importance and observing changes in model confidence.
Results: CAMAL demonstrated more faithful attention than Prior. Under pixel removal, CAMAL's model confidence dropped faster, indicating its attention correctly identified critical features. Under pixel insertion, confidence recovered faster. CAMAL achieved a 35.3% higher Area Under the Curve (AUC) than Prior in insertion scenarios.

Generalization Performance

DL: CAMAL consistently outperformed Vanilla in classification accuracy across most model-dataset combinations, with statistically significant gains in the majority of cases.
DRL: CAMAL showed a trend toward improved mean return compared to Vanilla, though the gains were not always statistically significant across all environments, suggesting attention supervision benefits generalization but may require refinement for complex DRL dynamics.
Comparison with Prior: CAMAL performed on par with or better than Prior in generalization, suggesting that the primary benefit of attention supervision lies in improving attention quality rather than drastically boosting task performance.

Significance and Claims

The paper claims that CAMAL demonstrates that spatial information in segmentation masks can be effectively leveraged to guide model attention. The authors emphasize that:

Trustworthiness is Key: Attention supervision must be grounded in trustworthy reference regions (segmentation masks) to be reliable; relying on pseudo-regions from zero-shot models leads to untrustworthy corrections.
Quality over Quantity: The primary benefit of attention supervision is the improvement of the quality of learned attention (alignment and faithfulness), with task performance gains being modest and context-dependent.
Scalability: The proposed batch-level extraction method makes attention supervision practical for modern, large-scale training without increasing inference costs.

The authors conclude that CAMAL establishes a reliable baseline for future attention supervision methods, highlighting that realizing the full potential of such methods requires grounding them in accurate, ground-truth spatial information.